void TFE_ContextSetSoftDevicePlacement(TFE_Context* ctx, unsigned char enable,
                                       TF_Status* status) {
  tensorflow::unwrap(ctx)->SetAllowSoftPlacement(enable);
}

void TFE_ContextSetLogDevicePlacement(TFE_Context* ctx, unsigned char enable,
                                      TF_Status* status) {
  tensorflow::unwrap(ctx)->SetLogDevicePlacement(enable);
}

// Enables running eager ops as function.
TF_CAPI_EXPORT void TFE_ContextSetRunEagerOpAsFunction(TFE_Context* ctx,
                                                       unsigned char enable,
                                                       TF_Status* status);

// Enables rewrite jit_compile functions.
TF_CAPI_EXPORT void TFE_ContextSetJitCompileRewrite(TFE_Context* ctx,

class GraphErrorInjectionPass : public tensorflow::GraphOptimizationPass {
 public:
  static bool enabled_;
  GraphErrorInjectionPass() {}

  tensorflow::Status Run(
      const tensorflow::GraphOptimizationPassOptions& options) override {
    if (!enabled_) {
      return absl::OkStatus();
    }
    if (first_call_) {
      first_call_ = false;
      return absl::OkStatus();
    }

  TF_SetConfig(&options->session_options, proto, proto_len, status);
}

void TFE_ContextOptionsSetAsync(TFE_ContextOptions* options,
                                unsigned char enable) {
  options->async = enable;
}

void TFE_ContextOptionsSetDevicePlacementPolicy(
    TFE_ContextOptions* options, TFE_ContextDevicePlacementPolicy policy) {
  options->device_placement_policy = policy;
}

  }

  Status AddInput(AbstractTensorHandle* input) override {
    GraphTensor* t = dyn_cast<GraphTensor>(input);
    if (!t) {
      return tensorflow::errors::InvalidArgument(
          "Unable to cast input to GraphTensor");
    }
    TF_AddInput(op_.get(), t->output_);
    return absl::OkStatus();
  }
  Status AddInputList(absl::Span<AbstractTensorHandle* const> inputs) override {

      config->mutable_graph_options()->mutable_rewrite_options();
  rewrite_options->set_scoped_allocator_optimization(
      tensorflow::RewriterConfig::ON);
  rewrite_options->mutable_scoped_allocator_opts()->add_enable_op(
      "CollectiveReduce");

  if ((tensorflow::IsGoogleCudaEnabled() || tensorflow::IsBuiltWithROCm()) &&
      num_virtual_gpus > 0) {
    tensorflow::GPUOptions* gpu_options =

unsigned char TF_SetXlaEnableLazyCompilation(unsigned char enable) {
  tensorflow::BuildXlaOpsPassFlags* flags =
      tensorflow::GetBuildXlaOpsPassFlags();
  bool original = flags->tf_xla_enable_lazy_compilation;
  flags->tf_xla_enable_lazy_compilation = enable;
  return original;
}

unsigned char TF_SetTfXlaCpuGlobalJit(unsigned char enable) {
  tensorflow::MarkForCompilationPassFlags* flags =
      tensorflow::GetMarkForCompilationPassFlags();

  fi
}

# The Python package is not allowed to depend on any CUDA packages.
@test "Pip package doesn't depend on CUDA" {
  bazel cquery \
    --experimental_cc_shared_library \
    --@local_config_cuda//:enable_cuda \
    "somepath(//tensorflow/tools/pip_package:wheel, " \
    "@local_config_cuda//cuda:cudart + "\
    "@local_config_cuda//cuda:cudart + "\
    "@local_config_cuda//cuda:cuda_driver + "\

    case DT_VARIANT:
      return true;
    case DT_QINT8:
    case DT_QINT16:
    case DT_QINT32:
    case DT_QUINT8:
    case DT_QUINT16:
      return tensorflow::flags::Global()
          .enable_quantized_dtypes_training.value();
    default:
      return false;
  }
}

template <typename Gradient, typename BackwardFunction, typename TapeTensor>

        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(
            TFE_NewExecutor(is_async, /*enable_streaming_enqueue=*/true,
                            /*in_flight_nodes_limit=*/in_flight_nodes_limit)),
        op_(nullptr),
        thread_(tensorflow::Env::Default()->StartThread(